Let's see if we have Kevin in the studio. Let's bring him in. He's here. Fantastic. How you doing?
Yeah, this is wild. I just like joined a Zoom and I'm live with you guys.
We're live. Yeah, it's great to have you. Welcome. I was watching you on X and thinking how weird it was that in 30 seconds I was going to be on the screen. This is awesome.
Yeah, we're leveraging this thing. It's called the internet and technology.
But we'd love to get you up to speed on it since it's a funny story. So Tyler Cowan called into the show before one of your guys' earlier releases and he was like, 'AGI is here.' Like he was basically saying like, 'It's a couple days out,' but he couldn't get his camera working. And it was this like funny dichotomy of like AGI's here, but like the internet is still actually...
There was actually another day when we had to basically take the whole show down because Zoom and both Zoom and Google Hangouts just went completely down like nationwide. And so we were like, well, we can't do our show now, I guess.
The running joke is that audio video is like ASI complete.
Yes. Well, everything else in the world is solved.
Yeah. Wait, so I mean I think long term for the show, we need to go proprietary. We need to build our own video streaming stack. Can you help with that? Codeex.
You know what can help with that? Yeah, code can help with that.
Break it down. How would I use Codeex? And by the way, you get to talk about AI and then you're going to have Blake from Boom here. So, you get to talk about supersonic flight afterwards. You guys have the coolest lives ever. It's full stack. Full stack technology. But yeah, can you take us through the announcement? Break down exactly what launched, when it's available, to who, what you're excited about, and then we'll go into some of the trade-offs in the product design and development.
Yeah, let's do it. So, we just launched this morning Codeex, which is a cloud-based software engineering agent that can work on many tasks in parallel. So a lot of folks are used to the kind of Cursor, Windsurf style, you know, GitHub Copilot style of AI development, which is really more about augmenting a single engineer, right? You're writing code in your IDE and you can press tab tab tab and it'll autocomplete and you're, you know, 10, 20, 40% faster. With Codeex, you actually have a software agent that runs in the cloud that can do entire tasks for you. So, you give it a task, it goes off and does it, and suddenly you have a PR and you didn't work on it at all. And it's powered by a version of o3 that we fine-tuned specifically to be really good at these kind of hard software tasks. We're super excited about it. It launched today inside ChatGPT. So, you can use it if you're... it's rolling out now to Pro, to enterprise, to Teams. Well, it'll make its way to Plus in the coming weeks. But the cool thing is it's an agent. It's a software agent in the cloud. So, you know, you're using it today from ChatGPT. You can imagine using it in the future from, you know, your terminal, your IDE, all kinds of other places and even, you know, hooking it up via API to your bug queues and just having this software agent churn through every single bug that you have. You know, for each one, look at the context of the bug, understand your codebase, and then suggest proactively how you would fix that bug and give you a PR that you can review. So, the world of software engineering is changing. We're super excited about it. It's a research preview. It's not perfect yet, but I think this is the future.
Yeah. So, I mean, I've already noticed that ChatGPT has been writing code for me for a while. It seems to be writing more and more code. I was telling Jordy about how I wanted to know the height of a desk in an image and I knew how tall a person next to it was. And I thought that it would just use images in ChatGPT to kind of one-shot this or guess. But it wound up spinning up and looking at the individual pixels with a bunch of Python code. I think it wound up writing like 500 lines of code and it got it really right. Unsatisfactory because it was just an average size table. It was like literally the standard size, but it really knew it. But so I'm wondering like is this something that will feel like a deep research project where I click a button to say hey let's use Codeex for this, I'm giving you the hint, or is this something that can be automatically triggered just from a text interaction like images in ChatGPT?
Yeah. So it's a... I mean by the way you know you never know if that desk was actually like 38.2 inches. So yeah it's worth it write the 500 lines of code. Yeah. Why not? You know, it's too cheap to meter. Actually, that's been one of the big... the coolest thing about o3. o3 personally for me has been a kind of feel the AGI sort of moment using that model. The things that it can do and a lot of it comes from the fact that it can use tools while it reasons. So it's thinking and in the process of thinking it can do some web searches and then it can take what it learned from a web search and write some code and then after that code it can do image analysis and then it can write some more code and do another web search and then finally with all of that context that will output the answer. It's like it's really been an unlock for a huge number of use cases and that's the kind of thing that enables Codeex here because you're right ChatGPT has been able to write code for a while right you can just go to ChatGPT and type in like here's a... write me code to sort this array of integers that I have and it can give you the code and it's going to be great at it. The difference here is Codeex is built to work on big complex code bases. So you're not just saying like do this little task for me. Write this function. You're saying I have a bug and I don't know where it is. There's a, you know, 100,000-line codebase. Can you please go understand my codebase and try and fix this bug? Or I'm a new engineer at a new job. I'm trying to understand what the heck is going on in this codebase. Can you explain to me where the code does X or Y or Z? And very quickly it'll look through the code and give you an explanation of how something works. You know, it's funny. I've even seen examples where people go, 'Hey, Codeex, find a bug in this codebase and fix it. Just find one randomly.' Just go do complex stuff and do hard work on a huge amount of existing context that differs from what you've been able to do in ChatGPT for a long time.
Yeah. Wild. From a personal perspective, I used to write Python pretty regularly. I haven't written much code, so I haven't really gotten into the Cursor, Windsurf world. Obviously I've been writing code via ChatGPT now. I noticed recently I kicked off a deep research report and it prompted me to put it on effectively a cron job. It was like, do you want me to just run this for you every week? And I said, yeah, that sounds awesome. That's great. But if I'm like a... I don't have a repo, but I could set one up. Is there a world where me as kind of like a prosumer non-technical user should set up a repo to house the custom code that Codeex writes for me to make a better experience for all the little custom software and tools and random stuff I use? Or should I just live in the o3 world where the code is pretty much ephemeral?
I think it depends what you're looking to do. You know, for simple things where you just want to like quickly put together a script or something, running it inside ChatGPT and not using version control and all of that is fine. But if you're going to do something, you know, that you expect to be longer lived, like it's the basis for something you actually want to build for yourself and maintain, then I think setting up a quick GitHub repo and using Codeex on it makes a ton of sense. By the way, I... so I used to be an engineer. I still dabble and screw around on the side and write code but nothing major and I hadn't written any code at OpenAI. I've been there about a year and haven't checked in a thing. Like Tuesday night I think. I was doing the rest of my work and I was like I want to fix a couple bugs and so I went and found a couple really basic bugs because I didn't want to screw anything up and sent Codeex off to work on both of them in parallel. Check back in a few minutes and I had two PRs. They looked right. So I submitted them, got them code reviewed by somebody, you know, who's actually a good engineer, and they were submitted, and now I've got a couple commits in the codebase.
So it's just... this is stolen valor, though. This is stolen valor. You didn't write that code.
But it's cool like I actually just kept, you know, other than just like getting to play around with the product and offering a little bit of feedback on a couple things, I was off doing the rest of my work and I had this software agent working for me in the cloud writing code.
Yeah. Talk about... I'm so curious to hear about the kind of internal testing process and when you guys decided the right time to actually roll this out as a research preview because I imagine you've been using Codeex in one form or another internally for like a very long time. Maybe it didn't have a name or anything like that, but I'm sure that ChatGPT has been contributing to the effectively the ChatGPT codebase almost since the beginning in some form or another.
Yeah, for sure. I mean, we're big users of our own tools. There was a version for a while that was mostly about the kind of how do you come up to speed quickly in a big codebase that was good at understanding our codebase and answering questions about it that was really popular with new engineers on the team and then once you kind of get to understand it you might not use that tool as often unless you're exploring a new area of the codebase. So that was like a proto version of this. But we've been working a lot over the last six months at improving the ability of our models to code. Like you've got GPT-4.1 which we released a little while ago which is very quickly become a really popular model. It's now I think default in Windsurf. It's increasingly a large percentage of Cursor users coding. And you know that came from focusing on the things that matter in creating a really good coding model that you can rely on means really good instruction following, longer context, you know, the ability to like not just make the changes but to make the changes the way a developer would. So don't add a bunch of extraneous stuff, don't add weird comments, you make surgical precise changes that accomplish the job. So there's a style element to writing good code not just a correctness element. And we've been focusing on all of this. And then you kind of bring that together with o3 and all of the things that... o3's ability to tool call and to reason and suddenly you can put together a really good coding model. So we've been thinking about this for a long time. This is the first time when we're like, 'Okay, this is now good enough that we think it deserves being a product for the rest of the world and we're excited to see how people use it.'
Can you talk a little bit about product design, product like inspiration in product design? I noticed like the very first iOS app had these incredible haptics when the tokens were streaming through that I hadn't really seen anyone do. I just opened the app last night and saw that when you're using voice mode to dictate to it, it has a different modal now. It feels like there's a very strong design language evolving. At the same time, there was, you know, people complaining, oh, I can't even... I'm using ChatGPT, I can't even stay logged in. That bug obviously got crushed pretty quickly. But what is the actual product design inspiration? Are there people that are pulling from certain schools of thought or anything or like is there someone driving that internally or is it just like baked into the culture?
Yeah, for sure. Ian Silber leads our design. I was fortunate to work with him at Instagram. He's an incredible designer. Building for ChatGPT is a really interesting thing because we have, you know, well over 500 million weekly active users at this point. So, it's a big scaled product. And so for products of that size, one of the most important things is to simplify, right? You're not just serving power users at that point. And so you're serving people that are just trying to get something done in their day. They don't care about the complexity. They don't care what the models are called. They just have a task and they want to complete it and you want to help them. But then on the other hand, we have folks who are super deep AI enthusiasts and want to, you know, digest every single new model that we use and try it out in all these different ways. And so we want both and we want them to... we don't want to sort of soften the edges of their experience. We want them to be able to do everything they possibly can to experience all of the power of AI. And so we both want to simplify the experience for a lot of our users and we want to provide the people that want it all of the bells and whistles. And so we try and balance that. You know, so we try and make it so that you don't need to worry about things like the model picker as much. You don't need to have a bunch of AI knowledge in the background to do what you want to do in OpenAI or in ChatGPT. But if you have that, you should be able to expose the sharp edges and test the different new features and stuff. And so we really actually try and get both of those things right. And it's a delicate balance.
